A 125B model on your gaming PC, plus three projects under a month old

The Lens

Strata runs Qwen3.8-Flash-Next, a very large open-weight model, on a gaming PC. The model has 125B parameters but only 6B do work on each token, so Strata keeps the busiest parts on your graphics card and the rest in system RAM, where the CPU handles them. It serves an OpenAI- and Anthropic-compatible API on localhost. The engine is MIT and free; the model weights carry Qwen's own community license. RAM is the real requirement, not the GPU. You need an NVIDIA RTX 20 to 50 series card or a recent AMD Radeon with 12 GB of VRAM or more, at least 32 GB of system RAM (enough for the coding variant only; 64 GB fits every size), and about 80 GB of free disk. On the author's RTX 5070 with 64 GB of RAM, a short chat writes roughly 50 to 95 tokens per second depending on model size. By default it answers one request at a time; you can raise that, but each answer gets slower on a 12 GB card. Solo developers with the hardware: a way to run a frontier-sized open model locally at $0 per token. For general local models, llama.cpp (ggml-org/llama.cpp) and Ollama (ollama/ollama) are the mature picks. The catch: the project is under three weeks old, one author writes nearly all of it, and engine releases land every day or two. The installer downloads a prebuilt engine and checks it against the SHA-256 GitHub publishes, which catches a tampered download but not a bad release, and it installs anyway if it can't get a hash. Cautious users should build from source (--build). Also read the setup prompts: the "experimental speed projection" it offers is, by its own docs, a vector that strips the model's refusals, not a speed boost.

The Lens

Coucou puts a little character in your MacBook's notch (or at the top of the screen on Windows and Linux) that watches your AI coding agents: which files Claude Code, Cursor, Codex or Gemini CLI are reading and editing, the diffs, and when they finish. You can answer permission prompts from the notch, from your iPhone's Lock Screen, or (paired with the Windows or Linux app) from an Android phone. The code is MIT and there's no paid tier. On a Mac, download the release, drag it to Applications, and click Install hooks; it backs up your Claude Code settings and shows the diff before writing anything. The Mac build is signed and notarized. The Windows installer isn't code-signed yet, so expect an "unknown publisher" SmartScreen warning the first time. Some agent integrations (Gemini CLI, OpenCode, Amp) are Mac-only. The iPhone app is a TestFlight beta, and the Android app has no signed release yet, so you build it from source. Solo devs running long agent sessions who want to step away from the keyboard get the most out of it. There's nothing for teams to buy or manage. If you only want a general notch utility, boring.notch (TheBoredTeam/boring.notch) is the established open source pick. The catch: approving agent actions from your phone is a real security surface, even behind Face ID or a fingerprint. It's mostly one developer shipping close to a release a day since launching in late September 2026, and the scope keeps widening (mail, crypto prices and Spotify pills are recent additions). And forks can't reuse the name, character, icon or sounds, because those assets aren't MIT.

The Lens

AIHOT is the full code behind aihot.news, a Chinese AI news site. It pulls items from RSS, web pages, APIs, X and WeChat accounts, has an LLM filter and score each one twice, writes Chinese headlines and summaries for what passes, and publishes daily, weekly and monthly digests. Swap in your own sources and scoring prompts and you get a news site for law, finance, or whatever you follow. MIT, free. Self-hosting is Docker Compose with five containers on a server with at least 2 cores and 4 GB of RAM. You need an OpenAI-compatible LLM key, and X or WeChat collection goes through paid per-request APIs. The author's own test run used about 930 model calls to import 152 demo items, so budget for tokens. The live site is free, and its RSS, API and MCP feeds need no key. Solo: read aihot.news, or deploy your own for a niche. Small teams: a cheap way to run an industry briefing without a human editor. The catch: everything is in Chinese, the docs and the output. It ships 18 demo feeds, not the live site's real source list, so the curation that makes aihot.news good is yours to rebuild. And the AIHOT name and logo aren't covered by the MIT license.

The Lens

Universal Modder turns Claude Code (or Codex, Gemini CLI, Cursor and others) into a game-modding agent. It bundles skills for engine fingerprinting, reverse-engineering playbooks, asset generation, in-game testing and packaging, plus a Python CLI that scans your Steam, Epic and Xbox installs and ranks the routes for modding each game. MIT licensed. Install it as a Claude Code plugin; you also need Git, Python 3.10+ and ffmpeg, plus Blender for 3D-to-sprite renders. Asset generation runs through fal's hosted models with a fal API key, or a local ComfyUI server for images without one. It asks before driving your mouse and keyboard, installing loaders into game folders, or publishing. The plugin is free; the art isn't. fal bills per generation, about $0.08 per image on the default editing model, with 3D, sound and voice priced separately, and the skill tells the agent to warn you before running more than about 20 generations. Solo hobbyists modding games they own are the audience. The author works at fal, which the README doesn't mention. The catch: the safety rules (no injecting into online games with anti-cheat, no cheats against other players, hands off DRM) are instructions to the agent, not enforcement. Its scope covers multiplayer games and servers you host, and the shared knowledge base already includes WoW private-server notes. It launched in late September 2026 with no versioned releases yet.